Decoding the nonstationary neural activity in motor cortex for brain machine interfaces

Abstract

Previous decoding algorithms used in brain machine interfaces (BMIs) usually seek a static functional mapping between the spatio-temporal neural activity and behavior and assume that the neural spike statistics do not change over time. However, recent work indicates the significant variance in neural activities, which suggests the nonfeasibility of the stationary assumptions on the neural signal sequences. To track the time-changing neural activity during the nonlinear decoding process, we developed a time-varying approach based on general regression neural network (GRNN) with a dynamic pattern layer. Applied on both simulated neural activity and in vivo BMI data extracted from rat's motor cortex, the proposed method reconstructs the movement signals better than the original GRNN algorithm with static pattern layer, which raises the promise of successfully tracking the time-varying neural activity for BMIs decoding. V

Long-term decoding stability of local field potentials from silicon arrays in primate motor cortex during a 2D center out task

“…The variation in preferred directions and the number of tuning directional signals shown in figure 5(b) in our study further confirmed the variability of neural encoding under the effect of nonstationarity of neural activities. Previous studies showed the non-stationarity could deteriorate the decoding performance of spike signals in static decoder over time even in several hours during the same session where no degradation of neural recording was observed [33,34]. Recently, Flint's study also found significantly reduced accuracy in offline decoding of LFPs for almost all sessions after day 0 compared to the accuracy on day 0, while the recording quality of neural signals did not obviously decay [22].…”

“…To reduce the effect of non-stationarity on long-term decoding stability, adaptive decoders have been used which could track the variable mapping between neural activities and kinematics. Several studies have proved that the offline decoding performance could be significantly improved by using adaptive decoders [33][34][35]. Therefore, it is a good way to solve the issues about the effect of non-stationarity on neural decoding stability.…”

“…Many previous studies suggested that the decline of decoding performance based on SUA over time was closely related to the degradation of SUA recording, which might result from biological responses, material changes and engineering failures occurring at the electrode recording site post-implantation [31,32]. In addition to the degradation of neural recording, the non-stationarity of neural activities across various behavioral conditions or attention states in chronic recording could also lead to fluctuations in decoding performance over time [33][34][35]. Although some studies demonstrated that the decoding performance of BMI systems could be improved by adaptive decoders in online experiments [34,[36][37][38], these studies were carried out in short-term periods when no significant signal degradation was observed.…”

Robust neural decoding by kernel regression with Siamese representation learning

“…One way to cope with the variability of neural activity is to use nonparametric kernel regression approaches [23][24][25][26]. Instead of learning a specific function, kernel regression approaches do not assume function forms and only assume that signals with similar inputs should share similar outputs.…”

“…Given incoming neural signals, they find the neighborhood of the neural data in the training set and compute the locally weighted average of the corresponding kinematics as the output [23]. In this way, kernel regression approaches can fit diverse relationships between neural signals and kinematics flexibly in subsets of data to model neural pattern changes; thus, more accurate and robust BMI performance can be achieved [24][25][26]. However, directly extending kernel regression approaches to neural decoding problems can be suboptimal.…”